Research Seminar in Mathematics - Towards Prioritized Policy Composition + Gaussian Processes in Machine Learning: Constrained Regression and Inverse Problems

Speakers

Finn Rietz and Jean-Paul Ivan, AASS, Örebro University.

Abstract (Finn Rietz)

In Reinforcement Learning, we want to learn behavior policies that solve a sequential problem of interest by maximization of the corresponding reward function. More specifically, we investigate how we can compose multiple such behavior policies, which is not straightforward in the general case, although highly desirable: Composable RL policies allow for knowledge transfer, data reuse, modular design, and zero-shot or few-shot adaption to more and more complex tasks. In this seminar talk, I will introduce our approach for prioritized policy composition, what makes it challenging, and show some preliminary results.

Abstract (Jean-Paul Ivan)

A perspective on Gaussian processes that is particularly useful in machine learning is the view that a Gaussian process specifies a Bayesian prior over a function space. The posterior distribution resulting from conditioning on observations of any linear functional on this space is available in closed form, leading to an interpretable and flexible approach to nonlinear regression problems. This talk will present this approach to regression and show how this it allows ill-posed inverse problems to be reframed as well-posed inference problems.

Welcome!